DEEP LEARNING-DRIVEN CELLULAR AUTOMATA FOR HIGH-RESOLUTION LAND USE CHANGE SIMULATION IN METROPOLITAN AREAS

Naba Kumar Rath, Sudhakar Sahoo

Abstract: Environmental monitoring, urban planning and disaster management are some of the most important areas where satellite image change detection has been really useful. The prescribed methodology deals with the major issues, complex pattern encoding and class imbalance in the remote sensing data. CNN is the effective for extracting the hierarchical spatial features and LSTM can model temporal dependencies between pair of images. These are featured with the integration of cellular automata that improves the local modeling of spatial relationships so that minute variations are detected better. The resulting architecture is novel in using the strengths of the individual components to produce a very robust and complete change detector. This approach is proved much more performant by extensive experiments that were performed on a difficult OSCD dataset. In the model, outstanding training and validation accuracies of 99.5 and 98.5, respectively, are obtained and the training loss falls to 18% after 100 epochs. The F1-scores are at 73% (training phase) and 70% (validation phase), which constitutes tremendous precision-recall balance. The check on the set of tests proves reasonable utility in application at 92.76 percent accurate and 24.17% loss and 63.49% F1-score, and generality is evident.

Keywords: Cellular Automata, Change Simulation, Satellite Imagery, LSTM, Deep Learning

DOI: 10.24874/PES08.02A.012

Recieved:   Revised:   Accepted:   
UDC:

Reads: 2   

---